Module 3: 2025–2026
What changed after classic prompting: agents, retrieval, structure, and measurement.
Models can call tools—search, code, calendars, browsers—and loop until a goal is met. Prompting shifts from one answer to planning, tool choice, and stop conditions.
Define allowed tools, success criteria, and what to do when uncertain. Prefer short plans over unbounded autonomy.
Retrieval-Augmented Generation pulls relevant docs into the prompt. Huge context windows help, but retrieval still wins for freshness, cost, and citeability.
Retrieve first, then instruct the model to answer only from provided sources—and say when evidence is missing.
Ask for JSON, tables, or schemas so downstream code can parse results. Many APIs now enforce JSON schema / constrained decoding.
Provide a schema example, required fields, and forbidden free text outside the structure.
System prompts set durable role, tone, safety, and output defaults. User prompts should carry the variable task and data.
Keep system prompts stable and versioned. Put volatile facts and documents in the user message or tool results.
Image, audio, and video inputs need the same clarity as text: what to notice, what to ignore, and the output form.
State the visual task explicitly (OCR, layout, defects, brand). Attach only the frames or crops that matter.
Treat prompts like products: build a small test set, score outputs, change one variable, and measure again.
Start with 10–30 golden examples. Track failure modes (format, facts, tone) separately from vibe checks.